Current Kidney Failure Mortality Prediction Models Miss the Mark for Clinical Use

New research is calling attention to the shortcomings of current mortality prediction models for people experiencing kidney failure, highlighting the need for new models with a lower risk of bias and better applicability in clinical practice.1

Results from the systematic review of 50 studies published between 2004 and 2024 suggest current mortality prediction models for this patient population are not sufficient for informing clinical decision-making and guiding treatment in people with kidney failure due to a high risk of bias and applicability concerns.1

According to the American Kidney Fund, an estimated 35.5 million people in the US have kidney disease and about 808,000 are living with kidney failure. Widely recognized as the fastest-growing noncommunicable disease in the US, kidney disease is often perceived as a “silent killer” due to a lack of signs or symptoms until the late stages in many patients.2

“Personalized risk predictions can inform the management of kidney failure for an individual and support decisions that best reflect their unique goals, preferences, and values. Yet, current guidelines do not include any recommendation to consult a mortality prediction model,” Pietro Ravani, MD, PhD, a professor and the Roy and Vi Baay Chair in kidney research at the University of Calgary, and colleagues wrote.1

To assess the quality and applicability of currently available mortality prediction models for people with kidney failure, investigators searched MEDLINE, Embase, and the Cochrane Library for studies published between January 1, 2004, the year Kidney Disease: Improving Global Outcomes (KDIGO) was originally established to develop and implement guidelines for the care of people with kidney disease, and September 30, 2024.1

The systematic review included studies creating mortality prediction models for people with kidney failure treated with long-term dialysis or people with a sustained estimated glomerular filtration rate (eGFR) < 15 mL/min/1.73 m2 with a prediction horizon of ≥ 3 months. Of note, investigators excluded studies not restricted to people with kidney failure; studies exclusively including kidney transplant recipients; studies limited to patients with acute kidney injury in hospital; and studies reporting associations rather than predictions.1

Investigators used the Prediction Model Risk of Bias Assessment Tool and followed prespecified questions about study design, prediction framework, modeling algorithm, performance evaluation, and model deployment.1

Of 7184 unique abstracts screened for eligibility, 77 were selected for full-text review. In total, 50 studies that created all-cause mortality prediction models were included, with 2,963,157 total participants with a median age of 64 (range, 52-81) years and a median proportion of women of 42% (range, 2%-54%).1

Investigators noted primary studies used traditional and/or machine learning methods for prediction model creation, including logistic regression (38%), Cox regression (42%), and random forests (2%) for survival outcomes. They also pointed out the most used measures of prediction performance were measures of discrimination ability, including the C index and the time-independent AUC (82%). A calibration plot was reported in 24 studies (48%).1

Further analysis revealed included studies were at high risk of bias due to inadequate selection of study population (54%), shortcomings in methods of measurement of predictors (30%) and outcome (24%), and flaws in the analysis strategy (100%). Investigators additionally expressed concerns regarding the applicability of the models, citing the fact that study participants (62%), predictors (34%), and outcomes (10%) did not fit the intended target clinical setting.1

Investigators acknowledged multiple limitations to these findings, including the included models’ potential utility for addressing research needs rather than informing clinical practice; the lack of people with kidney failure who still had to make treatment decisions or who chose conservative care without dialysis; and the fact that many studies were published before recommended reporting standards and quality assessment tools for risk prediction modeling were published.1

“According to this systematic review of 50 studies, published mortality prediction models for people with kidney failure were not found suitable to inform clinical decision-making,” investigators concluded.1 “We advocate for the use of existing guidelines and checklists to design, conduct, and report prediction modeling studies and the involvement of stakeholders in the study design to enhance model usability and clinical uptake.”

References

1. Jarrar F, Pasternak M, Harrison TG, et al. Mortality Risk Prediction Models for People With Kidney Failure: A Systematic Review. JAMA Netw Open. 2025;8(1):e2453190. doi:10.1001/jamanetworkopen.2024.53190
2. American Kidney Fund. Quick kidney disease facts and stats. July 19, 2024. Accessed January 6, 2025. https://www.kidneyfund.org/all-about-kidneys/quick-kidney-disease-facts-and-stats